The sentences your customer-facing AI agent must never say—and how to test for them
rcleffke · reddit · 2026-11-21
A practical guardrail playbook from a consultant deploying customer-facing agents: write a short list of forbidden sentences—unkeepable promises, off-list prices/discounts, and safety/legal/medical/financial claims—encode them as hard rules in the agent's instructions, define the fallback (hand off to a human), and gate every release with adversarial test questions including indirect phrasings like 'So you'll cover that, right?' Any failure blocks release until fixed; rerun the suite after every model, instruction, or document change.
More from coding & agent
- Agents can now buy stock market data autonomously via x402, from $0.02 per call — virattt · 2026-11-21
- "Claude already writes SQL better than me" — what's still worth learning — Zachly · 2026-11-21
- Claude already writes better SQL than me — so what's still worth learning? — Zachly · 2026-11-21
- Nous launches Hermes Marketplace with 2,514 plugins, including Claude-subscription provider — NousResearch · 2026-11-21
- Nous hackathon 8th place: Promise Keeper enforces plain-language rules on every tool call — NousResearch · 2026-11-21
- Nous hackathon 9th place: Hermes Canvas renders live agent plans driven by hooks, not the model — NousResearch · 2026-11-21